Most money spent on alternative data is wasted. The industry has ballooned to an estimated $11-19 billion in 2025 depending on which research firm you believe, with 94% of investment professionals planning to increase spending. The majority of that budget will buy data that is either too noisy to trade on, too widely distributed to contain alpha, or too legally fraught to use at scale.
That is the current state of alt data. A gold rush where most prospectors are digging in the same mine.
A Working Taxonomy
Alternative data falls into six broad categories. Not all are created equal.
Credit card transaction panels track consumer spending at the merchant level. Providers like Bloomberg Second Measure, Earnest Research, and YipitData aggregate anonymized transactions from panels of millions of U.S. consumers. The output is daily or weekly revenue estimates for publicly traded companies, broken down by brand, channel, and geography.
Web-scraped data includes job postings, pricing, product availability, and app store metrics. Thinknum is the dominant player here, tracking hiring trends, product listings, and pricing across thousands of company websites. The data is structured, timely, and covers a wide universe.
Satellite and geolocation data measures physical activity from orbit or from mobile device pings. Orbital Insight built its reputation on counting cars in retail parking lots and measuring oil storage tank fill levels using shadow analysis on floating roof tanks. Geolocation providers like Placer.ai and Advan track foot traffic via mobile SDKs.
Social media and news sentiment uses NLP to score the tone of text across Twitter, Reddit, news articles, and earnings call transcripts. Dataminr processes billions of public data points daily from over one million sources. RavenPack and MarketPsych Analytics turn unstructured text into numerical sentiment scores.
App usage and web traffic data measures digital engagement. SimilarWeb, Sensor Tower, and Apptopia track downloads, daily active users, session length, and engagement metrics across mobile apps and websites.
Expert networks and proprietary surveys provide qualitative or structured primary research. This is the oldest form of alt data, and it remains the least scalable.
Ranking by Predictive Value
Here is where I will be blunt. The hierarchy of actual predictive value, measured by out-of-sample information coefficients against subsequent earnings surprises, looks roughly like this:
Tier 1: Credit card transaction data. This is the gold standard. When you have a representative panel of consumer transactions, you are directly measuring revenue. Not a proxy. Not a correlation. Actual spend. Earnest Research launched its Spend Index (EASI) in late 2024 tracking spending across 89 merchant category codes and thousands of U.S. merchants. YipitData, which received $475 million from Carlyle in 2021, built its business on blending email receipt scraping with third-party card panel aggregation. Bloomberg Second Measure provides debit and credit card data to over 4,000 investors. The signal is strong because spending is the variable you actually care about. The main limitations are panel bias (U.S.-centric, skewed toward certain demographics, blind to foreign tourists and OTA bookings) and the 24-48 hour lag before transactions settle.
Tier 2: Web-scraped operational data. Job postings, pricing changes, product availability, and inventory levels are leading indicators of strategic direction and operational health. Thinknum tracks these signals across thousands of companies. The data requires more interpretation than transaction panels, but it captures dimensions that spending data misses entirely. A company aggressively hiring ML engineers six months before a product launch tells you something a credit card panel cannot. The legal risk is the main constraint. More on that below.
Tier 3: App usage and web traffic. Useful for consumer internet and SaaS companies where digital engagement is a direct proxy for revenue. Less useful outside that narrow universe. Signal quality degrades quickly for companies where digital channels represent only a fraction of total business.
Tier 4: Geolocation and satellite imagery. Overhyped outside of specific use cases. The two proven applications are retail parking lot traffic and crude oil floating roof tank monitoring. A Berkeley study found that parking lot fill rate data could generate 4-5% returns in the three-day window around quarterly earnings announcements. Orbital Insight's oil tank shadow analysis provides genuine informational advantage for commodity traders. Beyond these niches, the signal-to-noise ratio drops fast. Satellite revisit rates, cloud cover, and the difficulty of calibrating image data across seasons and geographies make this data expensive to maintain and hard to generalize. Funds that bought satellite imagery expecting a universal edge were mostly disappointed.
Tier 5: News and social sentiment. This category has experienced the most severe alpha decay. NLP sentiment analysis was a genuine edge in 2015-2018 when few funds were running it systematically. By 2025, the signal has been crowded into near-irrelevance for equity selection. Every major quant fund runs some version of it. GPT-class models have made the NLP itself trivially easy. When everyone can parse earnings calls and score tweet sentiment in real time, the informational advantage approaches zero. Dataminr still generates value for event-driven and risk management use cases, where speed matters more than exclusivity. But as a systematic alpha source for equity long/short, news sentiment is largely spent.
The Alpha Decay Problem
This is the central tension in alternative data. The moment a dataset becomes widely available, the edge it provides begins to erode. This is not theory. It is observable in the data.
A 2021 Refinitiv study found that hedge funds using consumer spending data saw a 10% improvement in quarterly stock prediction accuracy. That was when adoption was still climbing. By 2025, 67% of investment professionals report using alternative data. The predictive improvement has compressed as the same signals get priced in faster.
The dynamic is straightforward. Fund A discovers that credit card data predicts Walmart earnings with a two-week lead time. Fund A trades on it profitably for a few quarters. The data vendor sells to Fund B, C, D, and E. All five funds now trade the same signal. The market prices it in sooner. The lead time shrinks from two weeks to two days. The alpha decays from 200bps to 20bps.
Quandl, now Nasdaq Data Link after its 2018 acquisition, called alternative data "the deepest, least utilized alpha source in the world." That was true when Quandl was founded in 2011. It is less true each year. The platform now serves over 400,000 analysts with data from 350+ sources. When your "edge" is available to 400,000 other analysts, it is no longer an edge. It is a factor.
Greenwich Associates found that 42% of asset managers believe their alternative data edge lasts at least four years. I think that number is optimistic for most datasets. Exclusive proprietary data can retain value. Widely distributed panel data cannot.
Legal Risk Is Not Theoretical
The hiQ v. LinkedIn saga clarified some boundaries but left others murky. The Ninth Circuit ruled in 2022 that scraping publicly available data does not violate the Computer Fraud and Abuse Act. But hiQ itself was destroyed in the process. The company agreed to a permanent injunction, deleted all its data and source code, and paid $500,000 in damages after the district court found that creating fake LinkedIn profiles to access data crossed the line.
The practical takeaway: scraping public data is defensible under the CFAA. Violating terms of service, creating fake accounts, or scraping behind login walls is not. The GDPR adds another layer of complexity for any data involving European individuals.
In 2024, both Meta and Twitter lost lawsuits against Bright Data, reaffirming that scraping publicly available data remains legally permissible. But the cost of defending these suits is nontrivial, and the regulatory direction globally is toward more restriction, not less.
For alt data consumers, this means due diligence on data provenance is not optional. If your vendor cannot clearly explain where the data comes from and demonstrate compliance with applicable privacy laws, the legal liability transfers to you.
What Makes a Dataset Valuable
After spending years evaluating alternative datasets, three properties determine whether a dataset generates durable alpha:
Exclusivity. The number of funds with access is the single most important variable. If the answer is "anyone with a Bloomberg terminal," the data is priced in. The most valuable datasets come from proprietary collection methods or exclusive licensing arrangements. YipitData's early advantage came partly from having 70+ analysts covering just 70 tickers, providing a depth of interpretation that raw data feeds cannot match.
Granularity. Daily merchant-level transaction data is more valuable than weekly sector-level data. Store-level foot traffic beats MSA-level. The ability to decompose aggregate signals into company-specific, geography-specific, or product-specific components is where the information content lives.
Timeliness. A dataset that arrives two weeks before an earnings announcement is worth orders of magnitude more than one that arrives two days before. The value curve is nonlinear. Satellite imagery that updates weekly loses most of its value if competitors get daily feeds.
The intersection of all three is rare. Most alt data vendors prioritize breadth of coverage and ease of integration, which works against exclusivity. The funds that generate consistent returns from alternative data tend to either secure exclusive access to proprietary datasets or build internal data science teams that extract signals the vendor's off-the-shelf product does not surface.
Where This Goes
The alt data market will continue to grow. The $135 billion forecasts for 2030 may be inflated, but directionally the trend is clear. Institutional spending will increase because the cost of not having alternative data has risen. It is now table stakes for fundamental equity research.
But the returns to marginal alt data spending will diminish. The first dollar spent on credit card panel data generates substantial informational value. The tenth dollar spent on the fifth sentiment feed generates approximately none. The funds that outperform will be those that treat alternative data as a raw input requiring substantial internal processing, not a finished product they can plug into their models and expect alpha to appear.
The AIFI directory maps many of the companies mentioned in this piece under the Research and Data segment. The taxonomy is useful for understanding who does what. It will not tell you which dataset is worth buying. That requires the hard, unglamorous work of backtesting, panel bias correction, and signal decay monitoring that no vendor will do for you.